BLOGS

Every humanoid spec sheet leads with peak torque. The number that decides whether a joint survives its first hard footstrike — or can feel a fingertip contact at all — is one most buyers never see: the gear ratio, squared, multiplied by the rotor inertia. Engineers call it reflected inertia. This article unpacks the N² tax, why it splits the actuator world into two camps, and how to read a joint module's ratio like a disclosure document.

The N² Tax, Stated Plainly

A gearbox of ratio N multiplies torque by N. That part is on every datasheet. What the datasheet rarely says is that the same gearbox multiplies the motor rotor's inertia, as seen from the joint side, by N squared:

J_reflected = N² × J_rotor

The arithmetic is unforgiving. A 6:1 single-stage reduction reflects 36 times the rotor inertia. A 100:1 strain-wave stage reflects 10,000 times. A 160:1 arrangement reflects 25,600 times. Same rotor, radically different joint.

Why the square? Energy bookkeeping. For the joint to turn at one radian per second, the rotor must spin N radians per second, and kinetic energy scales with ω². Push on the joint and you are not accelerating a small rotor — you are accelerating a rotor that "wants" to spin N times faster than the motion you are imposing. The industry shorthand is that a high-ratio joint feels like a brick wall to an external impact: the limb cannot get out of the way of a collision fast enough, because before the controller has even detected the contact, the drivetrain has already presented the collision with an enormous apparent mass.

This is the N² tax. It is paid at every contact event, whether or not the control software is sophisticated enough to notice.

Why High Ratios Still Run the Industrial World

If reflected inertia is so costly, why do most industrial arms and many humanoids still ship 100:1-class transmissions? Because the tax only comes due at contact, and factory robots spend most of their lives avoiding it. Meanwhile high ratios collect dividends everywhere else:

• Holding torque is nearly free. A 100:1 stage lets a small motor hold a heavy static load with modest current. A low-ratio joint must burn continuous current — and continuous heat — to do the same. For an arm holding a pose for hours, the high-ratio design wins the thermal budget outright.
• Positioning is finer. Encoder resolution and control granularity are effectively divided by N. Near-zero-backlash strain-wave gearing makes repeatable, sub-arcminute positioning achievable in a compact, coaxial package — the default for precision manipulators for good reason.
• Slow, load-carrying gaits favor them. As the MIT proprioceptive-actuator work itself concluded, a load-carrying robot walking at slow speeds can energetically benefit from a higher gear ratio compared with an agile robot running at high speeds. The trade is application-dependent, not ideological.

The honest framing: high ratios buy stiffness, precision, and cheap static torque, and they charge for it in impact exposure and loss of transparency. Low ratios charge a different bill — current, heat, and motor mass — and pay back in survivability and feel.

The Impact Math: What a Footstrike Has to Move

The trade has been quantified. The "proprioceptive actuator" line of research (Wensing, Kim, and colleagues — IEEE Transactions on Robotics, 2017, MIT Cheetah lineage) introduced an impact mitigation factor to compare what different transmissions let a legged machine absorb. The comparison that matters most: a single-stage planetary reduction of about 5.8:1 against a conventional humanoid drivetrain built around roughly 160:1 reductions. Per the published model, the low-ratio design carried roughly 0.010 kg·m² of reflected inertia where the high-ratio humanoid carried about 0.085 kg·m² — on the order of eight times more — and across leg configurations the low-ratio joint retained about 90 percent of the impact mitigation of a hypothetical series-elastic version of itself, where the high-ratio humanoid retained about 52 percent.

Read that as an engineering statement, not a slogan: a well-executed low-ratio joint gets you most of the shock absorption of an explicit spring, without adding a spring.

The Mini Cheetah made the philosophy commodity. Twelve identical modules, each a hobby-grade outrunner with a 6:1 single-stage planetary delivering roughly 17 N·m peak — the platform that produced the first quadruped backflip in 2019, and whose recipe was copied so widely that low-ratio modules now power a large share of research quadrupeds and humanoid prototypes.

The humanoid statement of the same logic is the Berkeley Humanoid, a 0.85 m, 12-degree-of-freedom research biped presented at ICRA 2025. All four of its custom actuator types use the same 9:1 planetary reduction; the authors credit the small-ratio planetary choice with helping "reduce overall friction and minimize the quadratic amplification of cogging torque, friction, and reflected inertia errors." Its knee delivers 81.1 N·m peak and its ankle just 9.7 N·m: modest numbers by adult-humanoid standards, but predictable ones, which is exactly what a learning-based controller needs to cross the simulation-to-hardware gap. A predictable joint trains; an opaque joint simulates badly.

Transparency: When the Gearbox Is the Torque Sensor

Reflected inertia is only half of the low-ratio dividend. The other half is friction: gear friction, like inertia, scales brutally with ratio, and a low-friction joint lets the controller read contact torque directly from motor current at kilohertz rates. This is why a 2025 survey of humanoid locomotion and manipulation defines quasi-direct drive as a reduction below 10:1 and credits it with backdrivability and high force-control bandwidth from reduced backlash and friction — while warning of the flip side: high output torque demands high current, "which in turn causes strain on power electronics and overheating."

Two sensing consequences follow, and they reshape the whole sensor stack:

• In a QDD joint, current is the torque sensor. No joint-level force sensor, no elastic element. The measurement is implicit, high-rate, and free — but it inherits every drivetrain nonideality (cogging, ripple, friction) that the low ratio was chosen to minimize in the first place.
• In a high-ratio joint, current is nearly useless for torque. Friction through a 100:1 stage swamps the contact signal. The workaround is measuring output-side deflection across a compliant element with dual encoders, which restores torque sensing at the cost of extra components — and the deflection itself becomes part of the joint's loaded error budget, not a feature to ignore.

Buyers comparing joint modules are therefore really comparing two different sensing philosophies wearing the same housing. The question to ask is not "does it sense torque?" but "at what bandwidth, through what physical path, and with what calibration burden?" — and once torque transparency is settled, the next number to audit is torque bandwidth: how fast the closed joint can actually follow a torque command once the sensing path is in place.

Arms Join the Argument: Endpoint Inertia

Legs made the case for low ratios; arms are now extending it. A September 2026 arXiv paper (arXiv:2609.19194) describes AthenaZero, a bimanual research manipulator built from quasi-direct-drive actuation plus transmission remotization — actuators relocated toward the shoulder so distal links stay light. The claimed result: an effective endpoint mass comparable to a human arm, roughly an order of magnitude below conventional robot manipulators, with torque transparency inherited from the low-ratio drivetrain. On three baseball-inspired tasks the platform threw at speeds exceeding 30 m/s and caught and batted at speeds exceeding 14 m/s over a 7.3 m distance — dynamics that high-ratio arms, whose reflected inertia dominates their endpoint behavior, do not attempt.

The industrial implication is direct. Contact-rich manipulation — insertion, catching, handoffs, any task where the arm meets a moving object — punishes apparent endpoint mass exactly the way footstrike punishes leg inertia. As humanoids move from walking demos to physical work, the arm joints inherit the same N² accounting that legs have lived with since 2015.

The Per-Joint Ratio Map

No serious humanoid uses one ratio everywhere. Joints see different load regimes, and the N² tax lands differently on each:

Recent Chinese academic discussion lands on the same map. A Beijing Institute of Technology humanoid-systems lecture at the University of Chinese Academy of Sciences in late October 2026 described the modern mainstream as integrated rotary joints pairing low-speed, high-torque motors with small reduction ratios specifically to cut reflected inertia and ease force control — plus proximally-located drive units and foot-level buffering to shrink moving inertia and absorb landing shock. When a design philosophy shows up independently in MIT-lineage papers, ICRA bipeds, and Chinese university system reviews within the same year, it has stopped being a camp and become the consensus direction for dynamic joints.

This is also how Ti5ROBOT reads the market. Our Lightweight Joint Module line and Humanoid Robot Actuator line are organized as per-joint families rather than a single universal module, because a ratio optimized for a knee is a liability in an ankle — and because integrators assembling whole-body machines need a coherent torque-density ladder, not one number repeated six times.

A practical selection checklist for anyone auditing a joint module datasheet: (1) ask for reflected inertia at the output, not just rotor inertia and ratio separately; (2) ask for backdrive torque at rated speed — it bounds how physical contact will feel; (3) ask what impact test the joint has passed, in joules or in footstrike equivalents; (4) check whether the ratio differs per joint position or is one-size-fits-all; (5) ask for continuous torque at a static hold, thermally limited, not just peak; (6) if torque sensing matters, ask whether it comes from current, from dual-encoder deflection, or from a dedicated sensor — and at what bandwidth.

The Limits of Low Ratio

The N² tax is real, but so is the counter-bill. A low-ratio joint needs a physically larger motor for the same torque, mass that lands directly on the limb it drives; it holds static poses by burning current; and its power electronics must be sized for that current, not for average motion. The MIT-lineage conclusion remains the fairest sentence in the literature: the correct choice is application-dependent, since slow, load-carrying machines and agile, contact-heavy machines optimize for different points on the same curve.

Which is why the endgame is not a winner-take-all ratio. It is a per-joint allocation — high transparency where contact happens, high mechanical advantage where gravity holds court — with the reflected-inertia figure treated as a first-class spec alongside torque, bandwidth, and thermal limits.

The N² tax is unavoidable. Paying it knowingly is the difference between a humanoid that reacts to the world and one that the world has to walk around.

Source: humanoid.guide, "Quasi-Direct-Drive Actuators in Humanoid Robots" (2026-09-19)

Source: Gu et al., "Humanoid Locomotion and Manipulation: Current Progress and Challenges in Control, Planning, and Learning," arXiv:2501.02116 (2025)

Source: robo2u, "Humanoid Robot Hardware: The Ultimate Guide" (accessed 2026-11-02)

Source: Wensing et al., "Proprioceptive Actuator Design in the MIT Cheetah," IEEE T-RO 2017 (numbers via AI Wiki summary, cross-checked)

Source: arXiv:2609.19194, "AthenaZero: A low-inertia, bimanual robot for dynamic manipulation" (2026-09-15)

Source: AI Wiki, "Humanoid Robot" — Wensing et al. impact mitigation summary (accessed 2026-11-02)

Source: University of Chinese Academy of Sciences, BIT humanoid lecture report (2026-10-28)